Predicting social media addiction from Instagram profiles: A data mining approach
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Predicting social media addiction from
Instagram profiles: A data mining approach
Uso del data mining per predire la dipendenza da
Instagram in un campione italiano di studenti
Antonio Calcagnı̀, Veronica Cortellazzo, Francesca Guizzo, Paolo Girardi, Natale
Canale
Abstract In this short paper, we describe an application of data mining techniques
to predict Instagram users’ addiction from a set of features related to (i) Instagram
captions extracted from photos, videos, comments, and stories, and Instagram in-
dicators such as number of followers and following, blocked and closed friends,
and frequency of use. We first applied text mining to explore and describe the main
contents of Instagram captions. Next, we used a set of non parametric models and
ensemble methods to predict Instagram addiction as measured by the Instagram ad-
diction scale [1]. Models were compared via cross-validation using test and training
(random) sets from the original dataset. Results showed that Instagram addiction is
mainly predicted by the overall time spent on Instagram, writing stories and com-
ments, and number of followers. Moreover, the results suggest that Instagram users
made use of photos/videos and stories/comments differently, with the latter being
mostly related to emoticons, experiences, and relationships with other users.
Abstract Questo lavoro presenta alcuni risultati di una ricerca più ampia condotta
su un campione di giovani donne italiane circa l’utilizzo di Instagram come fa-
cilitatore dell’oggetivizzazione sessuale. In questo lavoro si presentano i risultati
relativi allo studio circa il modo con cui attività di Instagram come pubblicazioni
di foto, video, commenti e realizzazione di storie possano predire la dipendenza
dal social media, come misurato dall’apposita scala di Instagram addiction. Tale
Antonio Calcagnı̀,
University of Padova, e-mail: antonio.calcagni@unipd.it
Veronica Cortellazzo,
University of Padova, e-mail: veronica.cortellazzo@studenti.unipd.it
Francesca Guizzo,
University of Padova, e-mail: francesca.guizzo@unipd.it
Paolo Girardi,
University of Padova, e-mail: paolo.girardi@unipd.it
Natale Canale,
University of Padova, e-mail: natale.canale@unipd.it
1
534
2 A. Calcagnı̀, V. Cortellazzo, F. Guizzo, P. Girardi and N. Canale
analisi è stata realizzata mediante data mining, utilizzando una serie di modelli non
parametrici e metodi di ensemble. La validazione e la scelta del modello migliore è
stata effettuata via validazione incrociata. I risultati mettono in evidenza il ruolo del
tempo trascorso su Instagram, del numero di storie, commenti e followers come el-
ementi di predizione della dipendenza dal social network. I risultati, inoltre, hanno
altresı̀ evidenziato come gli utenti tendano ad utilizzare video e foto in maniera di-
versa da commenti e storie: questi ultimi, infatti, sembrano maggiormente connessi
a esperienze, emozioni e relazioni rispetto ai primi.
Key words: data mining, text mining, Instagram, social media addiction
1 Introduction
Instagram is a well-known social platform commonly used for personal reasons
as well as business activities. Over the last years, it has gained wide popularity
across the globe, becoming one of the most popular photo-sharing applications on
the Facebook platform [2]. In particular, recent trends show that Instagram is the
most important network being used among adolescents [3]. Recently, a number of
research have shown the role of Instagram in several psychological processes, such
as women objectification (e.g., see [4]). In this respect, appearance-related com-
ments on women’s bodies accompanying Instagram images seem to play a role in
body dissatisfaction as well as women self-objectification [7]. This and other results
suggest the importance of investigating the interplay between social media behav-
iors and social media addiction, especially in young users [5].
In this short paper we will focus on Instagram addiction in a sample of Italian
students. In particular, we will investigate the predictive role of Instagram contents
such as text (comments, hashtags, photo captions/descriptions), indicators (number
of followers/followings, blocked users, closed friends), and activity frequency on
social media addiction, as measured by the Instagram Addiction Scale [1]. Data min-
ing techniques have been used to analyse the data. Particularly, text mining was ap-
plied to textual components of the dataset (comments, emoticons, captions) whereas
a set of non parametric models and ensemble methods has been used to choose the
best model and predictors for the response variable Instagram addiction [6]. The
results suggest that the latter is mainly predicted by activities like interactions with
other users of the social networking system via messages, comments, and likes.
535
Predicting social media addiction from Instagram profiles: A data mining approach 3
2 Data and Methods
2.1 Data
The Instagram Addiction Scale [1] was administered to N = 97 female participants,
all of them using Instagram on a daily basis. Data were collected by using an online
survey.1 Subjects were between 18 and 31 years old with an average age of 23.64
years (standard deviation 2.23). They were asked to answer fifteen items grouped
into two sub-scales (i.e., Social effect, Compulsion) using an ordinal scale with six
anchors. As scales were standardized, final scores were computed by summing the
items corresponding to each sub-scale. Thus, the aggregated variable addiction was
then defined, with higher scores being indicative of higher levels of Instagram ad-
diction. For each participant, all Instagram data were also available in compressed
format (up to six months before). They consisted in comments to other posts, con-
nections with other users (followers, following, friends, blocked users), likes given
to media (photos and videos) as well as comments, media (photo, stories, videos),
searchers (texts, hashtags), and stories. For all these data, temporal information
regarding the use of the social network (time, day) were also available. The final
dataset comprises 94 subjects and 95 variables, including the response variable re-
garding Instagram addiction. Before running the analyses, 3 subjects were excluded
as they included missing observations (row-wise exclusion) whereas 36 variables
were left out from the analyses because of multicollinearity (as indicated by the
threshold r ≥ 0.9).
2.2 Methods
Data were analysed by means of data mining methods. With regards to the tex-
tual part of the data (photo captions, comments, hashtags), text mining descriptive
techniques (i.e., most frequent terms, bigrams, graphical analyses, sentiment analy-
sis) were used as implemented by the R library TextWiller [8] and tidytext
[9]. All texts from Instagram were pre-processed according to standard text-mining
pre-processing procedure (i.e., text-normalization, stopwords elimination, tokeniza-
tion) [10]. In order to predict Instagram addiction with respect to 95 features of
Instagram use, the following non-parametric and ensemble methods were instead
adopted: (i) Principal component regression (tuning: number of variables), (ii) Par-
tial least squares regression (tuning: number of variables), (iii) Lasso regression
(tuning: λ penalty parameter), (iv) Regression trees (tuning: α complexity parame-
ter), (v) Random forest (tuning: number of variables), (vi) Boosting (tuning: number
1 The institutional review board at the University of Padova gave ethical approval for the study
(protocol number: 2956)
5364 A. Calcagnı̀, V. Cortellazzo, F. Guizzo, P. Girardi and N. Canale
and depth of trees), (vii) Bagging. They were compared via 10-fold cross validation,
with prediction errors being computed via Out Of Bag (OOB) approach.2
3 Results and Discussion
Figure 1 shows wordclouds of the most frequent terms as extracted from captions
of Instagram photos, videos, comments, and stories. Results indicate how distri-
butions of photos/videos’ comments were different from those of stories: indeed,
words used in the first category were essentially descriptive (i.e., they just de-
scribed photos/videos in terms of reference to places or experiences like traveling)
whereas words used in stories/comments contained mainly emoticons and tag to
users. Therefore, Instagram users made use of stories/comments and photos/videos
differently. In general, sentiment analyses of captions revealed the presence of neu-
tral sentiments more frequently than negative and positive sentiments. In particular,
comments and stories showed a positive sentiment when compared to photos and
videos, which suggest that users tended to use photos/videos and comments/stories
with a different strategy.
(a) Photos and videos (b) Comments (c) Stories
Fig. 1: Wordcloud for captions of Photos, videos, comments, and stories.
In order to select the best subset of predictors of Instagram addiction from the
large set containing 95 features, we proceeded by running a set of non-parametric
models and ensemble methods as describe in section 2.2. Table 1 show the results
of model comparison on training and test sets, respectively. Among others, Random
2 More in details, ten train/test sets were created by bootstrapping the original dataset row-wise.
Then, models were estimated for each training data using a 10-fold cross validation; cv-errors were
computed and used to select the optimal tuned model. The best model was then chosen among those
having minimum RMSE, which was computed on the test sets using the out-of-bag observations.
Finally, the prediction error of the model was obtained as average of the ten RMSEs. Note that
cross-validation was used as implemented in the caret library [11].
537Predicting social media addiction from Instagram profiles: A data mining approach 5
forest achieved lower MAE and RMSE w.r.t. parameters estimation (training set)
and, in a similar way, it showed lower RMSE in the set as well. Thus, Random
forest was selected as the best predictive model of Instagram addiction.
RMSE Test set error
PCA regression 0.7769 0.8715
PLS regression 0.663 1.0411
LASSO regression 0.7189 1.0016
Regression Tree 0.7387 0.8615
Bagging 0.6017 0.7631
Random Forest 0.6403 0.7324
Boosting 0.6749 0.7909
Table 1: Model comparison on test set. Note that Test set error is computed via
bootstrap prediction error using 10 samples whereas the best model is represented
in gray tones.
Fig. 2: Random forest to predict Instagram addiction: (A) RMSE as a function of
the best number of predictors, (B) Variable importance plot for the final model. Note
that variable importance measures are computed via the importance() function
as implemented in the randomforests library [12].
The Random forest model was built on 500 trees which selected a final set of 67
predictors. Figure 2 shows the results of the final model. Overall, they suggested the
following main predictors for Instagram addiction (in order of importance): Over-
all time spent on Instagram (Tempo giorno insta), replies to users’ stories on
Saturday (Sab SAct), content searches during the night (Notte R), writing com-
5386 A. Calcagnı̀, V. Cortellazzo, F. Guizzo, P. Girardi and N. Canale
ments overall (num Commenti) or in a specific temporal window, namely on after-
noon (Pom cC) and on Wednesday (Mer C), number of followers (Followers).
4 Conclusion
We investigated the role of Instagram user’s data (captions, photos, comments, likes,
stories, hashtags, following and followers) in predicting social media addiction as
measured by Instagram Addiction Scale [1] in a sample of Italian students. We used
data mining techniques in order to (i) describe the use of Instagram commonly made
by users and (ii) find predictors of Instagram addiction from a large database con-
taining 95 features. The results suggested the importance of some variables, such
as time spent on Instagram, writing stories, comments, and number of followers, as
predictors of Instagram addiction. Moreover, the results also showed that users tend
to use photos/videos and stories/comments differently: the first were mainly adopted
for describing the objects they are related to. By contrast, the latter are more than
simple descriptions, as they include emoticons, reference to places and experiences,
tags to other users. Further studies need to be conducted to establish insights into the
various mechanisms of Instagram addiction. For instance, text mining results (e.g.,
emoticons, sentiments) may be used to define new predictors of social media addic-
tion, which would shed light on the role played by typed emotions on this emerging
phenomena.
References
1. Kircaburun, K., Griffiths, M.D., Journal of behavioral addictions 7(1), 158–170 (2018)
2. Li, Z., Agarwal, A., Management Science 63(10), 3438–3458 (2017)
3. Brown, R.C., Fischer, T., Goldwich, A.D., Keller, F., Young, R., Plener, P.L., Psychological
medicine 48(2), 337–346 (2018)
4. Meier, E.P., Gray, J., Cyberpsychology, Behavior, and Social Networking 17(4), 199–206
(2014)
5. Wallace, P., EMBO reports 15(1), 12–16 (2014)
6. Azzalini, A., Scarpa, B., Springer Science & Business Media (2009)
7. Tiggemann, M., Barbato, I., Body image 27, 61–66 (2018)
8. Solari, D., Sciandra, A., Finos, L.: Textwiller, Journal of Open Source Software 4(41), 1256
(2019). DOI 10.21105/joss.01256. URL https://doi.org/10.21105/joss.01256
9. Silge, J., Robinson, D.: tidytext, Journal of Open Source Software 1(3), 37 (2016)
10. Silge, J., Robinson, D.: Text mining with R: A tidy approach. ” O’Reilly Media, Inc.” (2017)
11. Kuhn, M., et al., Journal of statistical software 28(5), 1–26 (2008)
12. Liaw, A., Wiener, M., R News 2(3), 18–22 (2002). URL https://CRAN.R-
project.org/doc/Rnews/
539Structural entropy based modeling for
psychological measurement
Modeling Strutturale basato sull’entropia per le misure
in Psicologia
Enrico Ciavolino, Mario Angelelli, Paola Pasca and Omar Carlo Gioacchino Gelo
Abstract The contribution introduces the Entropy Based Structural Models and
their dynamical evolution: the Streaming entropy. The new variant is described and
tested on a typical clinical psychology scenario, that is, the psychotherapy process,
as no previous work made us of behavioral data to study the psychotherapeutic
relationship. Textual data consist of a psychotherapy transcript, whose word blocks
have been processed and classified according to their valence: positive, negative and
abstract. At first, the Stre-GCE algorithm computes one model parameter for each
interaction components, then parameters get updated in an alternate way (therapist-
patient, patient-therapist and so forth). Results show that Stre-GCE accounts for the
fluctuating nature of the psychotherapy interaction.
Abstract Il contributo introduce i Modelli Strutturali basati sull’Entropia ed una
loro evoluzione dinamica: la Streaming Entropy. La nuova variante viene de-
scritta e testata in uno scenario tipico della psicologia clinica, il processo psi-
coterapeutico, in quanto non esiste ancora un lavoro che utilizzi dati comporta-
mentali per esaminare la relazione psicoterapeutica. I dati testuali consistono nel
trascritto di un’intera psicoterapia, i cui blocchi di parole sono stati classificati in
categorie: positive, negative, astratte. Inizialmente, l’algoritmo Stre-GCE stima i
parametri di un modello per ciascuno dei componenti dell’interazione terapeutica,
per poi aggiornarli in modo alterno con riferimento all’interlocutore precedente
(terapeuta-paziente, paziente-terapeuta, ecc..). I risultati riflettono le fluttuazioni
tipiche dell’interazione psicoterapeutica.
Enrico Ciavolino
University of Salento, e-mail: enrico.ciavolino@unisalento.it
Mario Angelelli
University of Salento and INFN Lecce e-mail: mario.angelelli@unisalento.it
Paola Pasca
University of Salento e-mail: paola.pasca@unisalento.it
Omar Carlo Gioacchino Gelo
University of Salento e-mail: omar.gelo@unisalento.it
1
540You can also read
